[MLU] feat: add mlu support

[uqyxx]

This pr is for supporting Cambricon MLU backend

[gemini-code-assist]

Code Review

This pull request registers and implements a new CNCLCheckpointEngine for collective communication checkpointing on MLU hardware, and updates MLU platform configurations to map MLU resources to GPU resources in Ray. Feedback on the new checkpoint engine highlights several critical issues: blocking ZeroMQ operations within asynchronous methods, potential runtime errors due to missing attributes on StatelessProcessGroup during finalization, uninitialized instance variables, and premature garbage collection of local ZeroMQ contexts. Additionally, improvements were suggested to handle non-contiguous weights, align buffer offsets for mixed-precision dtypes, and optimize synchronization overhead by replacing global device synchronization with stream-level synchronization.

[heavyrain-lzy]

Hi @uqyxx ~ Thank you for contributing to this rep. We have merged the basic plugin method in the verl. Please update this PR according to the new contributor document. You can contact me if you have any questions.

[heavyrain-lzy]

Hi @uqyxx, we have added FlagOS engines for multi-chips that have been verified on the Nvidia platform. You can reference this #2 to add platform, engines, usr-guide, etc.

[heavyrain-lzy]

Hi @uqyxx, we have added FlagOS engines for multi-chips that have been verified on the Nvidia platform. You can reference this #2 to add platform, engines, usr-guide, etc.

@uqyxx Please solve the conflicts and add a user guide in the docs.

[uqyxx]

Hi @uqyxx, we have added FlagOS engines for multi-chips that have been verified on the Nvidia platform. You can reference this #2 to add platform, engines, usr-guide, etc.

@uqyxx Please solve the conflicts and add a user guide in the docs.

Hi @heavyrain-lzy , all done! Thx!

[heavyrain-lzy]

Hi,@uqyxx, have you verified that the example executes correctly?

[uqyxx]

Hi,@uqyxx, have you verified that the example executes correctly?

Yes I have tested several scripts in verl examples, can run succesfully

[heavyrain-lzy]

Hi @wuxibin89 , the PR has been reviewed and is ready to be merged. Do you have any comments from your perspective?

[heavyrain-lzy]

LGTM

[heavyrain-lzy]

@uqyxx Hi, I have added end-to-end validation coverage in #5, including E2E checks. Please follow the scripts: https://github.com/verl-project/verl-hardware-plugin/blob/main/scripts/baseline_grpo_gsm8k.sh and compare the result in the https://swanlab.cn/@heavyrain/verl_grpo_gsm8k_math/runs/8h196r8o/chart